TL;DR: llama.cpp released an update fixing a bug in SYCL's handling of reordered quantized weights, improving performance and reliability for specific hardware.
Summary: ggerganov/llama.cpp released version b11748, which includes a fix for the get_rows function on reordered quantized weights when using SYCL. This addresses a specific technical issue related to how quantized model weights are processed on SYCL-compatible hardware, ensuring correct data retrieval.
Why it matters: This fix is crucial for developers leveraging llama.cpp with SYCL for local inference, as it enhances the stability and accuracy of quantized model execution. Builders should update to this version to ensure optimal performance and avoid potential issues with quantized models on SYCL-enabled devices.
Source: github_releases